Evidence for Gradient Input Features from Sino-Japanese Compound Accent

要旨

This paper argues that pitch accent patterns of two-member Sino-Japanese compounds, hitherto considered unpredictable, can be strongly predicted by positing gradiently-valued accent features in the input, in the framework of Gradient Symbolic Computation "GSC", (Smolensky and Goldrick 2016). A simple machine-learning algorithm finds accent-affecting propensities = activations that collectively work for a set of compounds with frequently-occurring morphemes from the NHK corpus. I show that gradient input representations are needed to explain these kinds of phenomena. Examining a set of examples in which switches of morpheme order can change the accent pattern in ways that prosody cannot account for, I show that such phenomena can be explained by GSC but not by systems that have discrete-valued inputs and weighted, lexically-indexed constraints, thus providing evidence in favour of the GSC framework.

主題